摘要
针对说话人自动识别系统的性能与稳定性在高噪声环境下会严重下降,人耳却能捕捉高噪声环境中的目标语音的问题。提出使用能模拟耳蜗听觉特性的GFCC(Gammatone Frequency Cepstral Coefficient)特征与主成分分析(Principal Component Analysis,PCA)相结合的方法,以提高识别系统的鲁棒性。在不同程度信噪比的真实语音案件噪声条件下,对国际上最认可的基于似然比证据评估体系的法庭自动说话人识别系统的准确性和稳定性进行测试。实验结果显示:GFCC特征在多个程度的信噪比条件下,甚至信噪比为-20 dB的条件下,依然能保持较高的识别准确度和良好的稳定性,并能够提供可量化、可重复的证据强度值。
The performance and stability of speaker automatic recognition system are seriously degraded in high noise environment, but the human ear can capture target speech. In order to improve the robustness of the recognition system, we proposed a combination of gammatone frequency cepstral coefficients(GFCC) features and principal component analysis(PCA), which could simulate the cochlear auditory characteristics. The accuracy and stability of the most internationally recognized court automatic speaker recognition system based on likelihood ratio evidences evaluation system were tested under the conditions of different SNR of real voice cases. The experimental results show that the GFCC features can maintain high recognition accuracy and good stability, and can provide quantifiable and repeatable evidence strength values under various SNR conditions, even under the condition of SNR -20 dB.
作者
王华朋
姜囡
晁亚东
刘恩
Wang Huapeng;Jiang Nan;Chao Yadong;Liu En(Department of Video and Audio Materior Examination,Criminal Investigation Police University of China,Shenyang 110854,Liaoning,China)
出处
《计算机应用与软件》
北大核心
2019年第7期65-68,98,共5页
Computer Applications and Software
基金
2016国家社会科学基金重点项目(16AYY015)
辽宁省重点研发计划项目(2017231006,2017231004)
公安部公安理论及软科学项目(2017LLYJXJXY040)
关键词
GFCC
似然比
证据强度
科学证据
PCA
GFCC
Likelihood ratio
Evidence strength
Scientific evidence
PCA
